Repository navigation
DEP: Disallow np.promote_types(string, other) -> string #15897
Description
It appears that np.nan gets converted to the string "nan" when initializing an array.
Reproducing code example:
import numpy as np
np.array(["some string", np.nan])prints
array(['some string', 'nan'], dtype='<U11')
Similar behavior can be observed for other generators, such as hstack, concat, repeat and tile.
Expected behavior
I would expect that the above returns
array(['some string', nan], dtype='object')
with the second element of the array being the object np.nan.
Numpy/Python version information:
1.18.1 3.8.2
Reacted by Luca Bittarello
Activity
seberg commentedon Apr 2, 2020
Right
np.promote_types()for string and all other NumPy types returns strings, I agree with that being incorrect behaviour.objectwould be better, although I personally think we should consider deprecating it entirely (forcing you to writedtype=object, or maybedtype=np.allow_objectif that need arises).Note that this is a fairly large change maybe, so it would have to be very slow in either case. However, I think we pretty much have to do this change because whenever
np.equalknows about strings, it would be good if it is not an exception sincenp.equal(string_array, integer_array)for example must never return true (by casting the integer array to strings). The discrepancy would not be nice IMO.I suppose I did not open an issue for this specifically, so maybe we can use this for discussion.
[-]np.array and other generators convert np.nan to string "nan"[/-][+]DEP: Disallow `np.promote_types(string, other) -> string`[/+]seberg commentedon Apr 7, 2020
Hmmm, I do not actually remember what I thought was the big issue with the approach that I had explored, there is some slightly ugly part around ufuncs, but it seems not super bad. See gh-15925 for an exploration of this.
The main test will probably be how pandas handles this.
seberg commentedon Sep 23, 2026
There was a new issue at #32765, so I'll randomly close this one (althou, I guess this may at this time even have a bit more info, but oh well).